Apply transformation on SR?

[ChefLiutao]

In your implementation, a 7x7 sampling grid is output by LocalNet and then applied on SR results instead of HR images (Ground Truth). It is slightly different from your paper. Moreover, sampling 7x7 grid from 9x9 HR piece seems to be better than from 7x7 SR pieces?

[yhjo09]

Yes it is, and I mentioned this in the paper (p5). Because, if apply LocNet to GT then we get a slightly transformed GT and using this for training target deteriorates the PSNR of SR outputs. Assume we sample 7x7 patch from 7x7 image and the transformation matrix is [1 0 1, 0 1 0, 0 0 1] (x+1 pixel translation), then the rightmost column of the resulting 7x7 patch is filled with meaningless data (eg. zeros), so we have to sample from a larger size image.